Software Engineer, Applied AI

What matters most

• Build for real-world audio

  Architect and ship agents and an audio intelligence pipeline for noisy, unstructured offline conversations—not just clean online-meeting audio.

• Deliver natural voice interaction

  Build a voice-first interface that lets users command Rilla through speech, alongside user-facing chat and real-time inference.

• Make conversations searchable

  Develop a search engine that surfaces business-critical insights from voice data that has not previously been searchable.

• Own production AI foundations

  As an early AI engineer, shape techniques, tooling, and models across data acquisition, deployment, evaluation, and inference. The posting requires production AI/LLM experience and familiarity with eval frameworks, agent tooling, and prompt engineering.

• Learn directly from customers

  Talk with and visit customers in the field to understand practical needs. The posting emphasizes empathy for customers who work outside tech companies and solving real pains rather than vanity metrics.

Likely challenges to think through

• Audio quality versus useful insights

  Preparation hypothesis: Noise and unstructured speech may make extracted insights unreliable. Prepare to explain how you would evaluate an audio pipeline on representative field recordings, inspect failures, and decide when an insight is trustworthy enough to surface.

• Fast voice responses versus accuracy

  Preparation hypothesis: A natural speech interface may require low latency while agents need time to interpret audio and choose actions. Prepare to discuss latency budgets, error handling, and when to ask for clarification rather than act on uncertain input.

• Shipping quickly versus building foundations

  Preparation hypothesis: Early ownership across the AI lifecycle and frequent customer visits may compete with time for durable tooling. Prepare an example of using customer feedback to prioritize a production release while leaving a clear path for evaluation and iteration.

Topics to prepare

• Production AI/LLM lifecycle

  Required: experience building and deploying AI/LLM systems in production. Revise by diagramming a system from data acquisition through inference and a user-facing interface, then identifying how you would detect and address failures.

• Evaluation, agents and prompts

  Required: familiarity with evaluation frameworks, agent tooling and prompt engineering. Revise by defining test cases for an audio-based agent, including an ambiguous command and an unsupported answer, and explaining how you would compare prompt changes.

• Real-world audio intelligence

  The role involves agents and pipelines for noisy, unstructured offline conversations. Revise by sketching how audio quality and conversational ambiguity could affect extracted insights, with one check for each failure mode. The checks are preparation suggestions, not specified requirements.

• Voice-data search

  Rilla is building search for business insights in voice data. Revise by tracing a spoken customer question to a relevant conversation and a verifiable answer; discuss what should happen when the evidence is weak. This design exercise is a suggestion, not a stated implementation.

• Real-time voice interaction

  The posting describes a voice-first interface, real-time inference and user-facing chat. Revise by outlining the latency and error trade-offs in responding to a spoken command; no particular latency target is stated.

Skills to revise

• Build an API-backed AI interaction

  Python, TypeScript and FastAPI are listed technologies. As practice, implement a small FastAPI endpoint and TypeScript caller that submit a question and return a structured response; this exercises the interface work without implying a prescribed Rilla design.

• Run a focused AI evaluation

  The posting explicitly calls for familiarity with eval frameworks and prompt engineering. Create a small set of conversation questions with expected outcomes, run two prompt versions, and record answer quality and failure cases.

• Prototype an audio-processing step

  The role targets messy real-world audio, and PyTorch is listed among its technologies. As optional practice, use a short noisy sample to test one audio-processing step and inspect where it fails; no specific model or technique is required by the posting.

• Explain a real-time data path

  LiveKit, AWS, PostgreSQL, Redis and S3 appear in the technology list. Sketch a plausible path from incoming audio to an insight shown to a user, labeling where each technology might fit and which choices need validation. That allocation is a preparation exercise, not Rilla’s stated architecture.

• Practice customer discovery

  Working directly with customers, including field visits, is explicit in the posting. Draft five questions about how an offline sales conversation becomes a coaching decision, then turn a hypothetical answer into a testable product need rather than an assumed feature.

• Prepare a production AI project walkthrough

  The posting requires experience building and deploying AI/LLM systems. Prepare one example covering the problem, architecture, your decisions, deployment, measured results, and a failure you learned from; use only work you actually did.

• Sketch an offline-audio system design

  Rilla describes noisy, unstructured conversations, voice commands, and searchable insights. As a preparation exercise, draw a data-acquisition-to-inference design and explain how you would handle noise, latency, and unreliable outputs.

• Bring an evaluation plan

  The posting calls for familiarity with eval frameworks, agent tooling, and prompt engineering. Prepare a small test set and a comparison plan for prompts or agent behavior, including success measures and failure cases.

• Map your technical examples to the listed stack

  Rilla says it primarily uses Python, TypeScript, and FastAPI, alongside AI, real-time, cloud, and storage tools. Note which tools you have used, what you built with them, and where you would need to learn; do not present listed tools as individual requirements.

• Rehearse a customer discovery example

  Direct customer work and field visits are explicit parts of the role. Prepare an actual example of uncovering a practical need, changing a technical decision after feedback, and explaining the outcome without relying on vanity metrics.

• Draft questions about the role’s constraints

  Ask how the team evaluates noisy-audio performance, prioritizes voice commands versus search, and gathers field feedback. Also clarify the stated NYC in-person schedule of approximately 60 hours per week so you understand the working arrangement.